AI small business SME build vs buy buying software

Build vs buy for AI: when off-the-shelf is genuinely fine

We build custom AI systems for a living. Most of the AI in your business should not be one of them.

That isn’t modesty. It’s what the numbers say. MIT’s NANDA initiative looked at 300 public AI deployments alongside interviews with business leaders and found 95% of generative AI pilots produced no measurable return. The part of that study nobody quotes is what happened when you split the projects by how they were sourced. Work done with outside specialists landed about two-thirds of the time. Systems a company built internally, with its own people, landed about a third of the time.

Read the categories carefully before anyone waves that at you. “Build” there means a business deciding to become a software business on the side. Bringing in an outside specialist sat on the buy side of the line, which is worth knowing when a supplier quotes the stat at you as proof you should hire them, and worth knowing when we quote it at you too.

So here’s the version of this decision that’s actually useful for a business your size: which parts of your AI you should rent, which you should own, what both cost once you count properly, and the four questions that sort almost every process into one column or the other. If you’re earlier than this and still working out what you’re buying at all, what an AI audit is and how to choose an AI consultant come first.

The question is never build or buy. It’s which parts.

People treat this as one big decision. Stay on subscriptions, or commission something custom. All or nothing.

It’s never all or nothing, and it never has been. Your business is a list of processes, and the right answer differs for each one.

You’re not building a payment processor. Stripe exists. You’re not training a language model either, which is the AI equivalent, and anyone suggesting you should is selling you a science project. Those are commodities and you rent them forever.

Now take the thing your business does that nobody else does quite the same way. How you quote. How you triage what comes in. The order your team works through a job, and the six things they check that aren’t written down anywhere. That part isn’t a commodity, and off-the-shelf software has to serve a plumbing firm and a recruitment agency and you from the same template.

There’s a clean test for this borrowed from a build vs buy framework that cuts through the noise: is this software a commodity, or is it part of how you win? Commodity, rent it. How you win, that’s the part worth owning.

Most businesses have one or two processes on the second list. Not twenty.

What “buy” actually means for AI in 2026

Worth being specific, because “buy” now covers three different things at three different prices.

The AI already inside software you pay for. Your CRM, your accounting package, your helpdesk, your document storage. Most of them shipped AI features in the last two years and a good number of those features are included or cost very little. This is the cheapest AI you will ever get and it’s the layer owners most often skip, usually because nobody has been through the settings since the account was opened.

Seats in a general assistant. Microsoft 365 Copilot lists at £23.10 per user per month paid yearly on top of what you already pay Microsoft. For 20 people that’s £5,544 a year. ChatGPT and Claude seats sit in similar territory. Good value for the people who live in documents and email. No help at all for a process that runs across three systems.

A point tool that does one job. Transcription, invoice capture, call summaries, first-line email answers. Somebody sells nearly every one of these already, they’re cheap, and they work.

Owners underestimate how much of that second layer they’ve already bought. Zylo’s 2026 SaaS Management Index found ChatGPT is now the single most expensed application anywhere, with expense-based software spend up 267% year on year. AI arrived in most businesses on staff expense claims, not through a decision. You’ve probably already bought more of it than you think, and your team is running more of it than you’ve bought. Shadow AI covers how to find out what’s actually in use before you decide anything.

Buy, and keep buying, when any of these is true

Not a scoring system. If one of these describes your process, rent it and spend the attention somewhere else.

  • Somebody already sells this exact thing. Meeting notes, receipt scanning, appointment reminders, a chatbot answering your 30 standard questions. These are solved. Paying for a custom version is paying twice.
  • It costs under a couple of hours a week. Multiply the times it happens by how long each one takes. Under two hours and nothing you build will ever pay for itself, whatever it costs. The manual work cost calculator gives you that figure in about five minutes.
  • The process changed in the last six months. A build encodes how you work today. If how you work is still moving, you’ll be paying to rebuild it. Which processes to automate first covers how to tell when something has actually settled.
  • It lives inside one system. If the work starts and finishes in your CRM, the answer is almost always a setting in your CRM.
  • Nobody has counted it yet. This one is absolute. No number, no build. That’s the single biggest reason AI pilots never ship: everyone agreed it felt faster and nobody could say by how much when the renewal came round.

The UK picture backs up going slowly here. The British Chambers of Commerce found 54% of UK SMEs actively using AI in 2026, up from 35% a year earlier, and Forrester expects fewer than 15% of firms to switch on the agentic features in their automation platforms this year. Almost everybody is at the rented layer. Very few have built anything, and most of them are right not to have.

Build when the shape of your business is the problem

The tell isn’t ambition. It’s friction you’ve stopped noticing.

You’re paying for software, and then paying people to work around the software. Somebody exports a report on Monday so somebody else can re-key it on Tuesday. Everyone calls that “just how it works here”.

Gartner found 68% of fast-growing businesses regret a software purchase, and 31% have replaced software specifically because they couldn’t customise it to fit how they work. That second number is the whole argument. You don’t outgrow your tools because they’re bad. You outgrow them because they can’t bend, and one day the bending is being done by your team instead.

Four conditions. You want at least three before building is a sensible conversation.

  • The process crosses three or more systems that don’t talk properly. This is the commonest one by a distance, and it’s the one no subscription fixes, because every vendor’s AI only sees its own data.
  • The exceptions are the expensive part. Rules handle 70% of cases. The other 30% land on somebody’s desk scattered through the week, which is exactly why nobody has ever added them up. If the judgement calls are eating an afternoon a week, that afternoon is the business case.
  • The bit costing you money is the bit that makes you different. If you could buy it, a competitor could buy the same thing tomorrow. Owning the thing that’s genuinely yours is worth something beyond the hours.
  • The volume is real and it’s settled. Same process, same shape, high enough frequency that a percentage of it matters. Founderise is the clean example: the delivery process was proven and repetitive before anything got built, which is why it came out at 12 hours saved a week and a 3.5x margin increase rather than a pilot nobody renewed.

Both sides hide their costs. In different places.

This is where most build vs buy arguments go wrong, because each side compares its real cost against the other side’s advertised one.

Renting hides its cost in sprawl. It starts cheap and climbs forever. Every gap gets patched with another subscription, and businesses underestimate their own software usage by about 40%: for every 10 tools you think you’re running, there are 14. Then 52.7% of SaaS licences sit unused on Zylo’s data. You’re paying per seat, per month, forever, for a fair amount of nothing.

AI makes that worse rather than better. Flexera’s 2026 State of ITAM report found only 31% of organisations have accurate visibility into their AI software spend, and 59% say wasted AI spend has risen year on year.

Owning hides its cost in upkeep. The build price is the visible number and it’s not the whole number. Something has to pay for keeping the system current as the business moves underneath it, and whether you budget for that with your supplier or absorb it internally, you’re paying it. Anyone quoting you a build with no line for upkeep is quoting you something that works for a year. Third-party running costs for a small business system typically land somewhere around £150 to £400 a month, and an AI system pointed at a high-volume process sits at the top of that or above it, because it costs something every time it thinks.

Ignore that and you become a statistic. Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027, naming escalating costs first. KPMG found 49% of leaders have scaled back, narrowed, delayed, or paused agent deployments once expected costs outran expected value.

And one option that’s hiding nothing at all, because it’s simply out of reach: hiring engineers. A small in-house team runs well over £1m a year fully loaded before it ships anything. No business under 50 people should consider it. When we say build, we mean an outside team builds it, hands it over, and you own it. That last phrase carries more than most people assume, and who owns the code and the data sets out what it should cover. What AI actually costs a small business has the full arithmetic on both paths.

Four questions that settle it

Run one process through these. The answers usually point the same way.

1. Could you buy this exact thing today? Not something like it. This. If yes, buy it and move on. The pile of AI point tools on the market is deep and most of them are cheap.

2. Is the part costing you money also the part that makes you different? If a competitor could buy the same subscription and get the same result, it’s a commodity. Rent it.

3. Has it been the same process for six months? Still moving means still too early. Settle it on cheap tools first.

4. What does it cost you by hand, per week, in pounds? If you can’t answer this one in numbers, nothing else on the list matters yet. Go and count.

If you’d rather run the whole business through this than one process at a time, the systemisation scorecard ranks them for you and tells you plainly where the honest answer is to leave something alone.

The answer that’s right most often is both

Real systems are layered, and the good ones are mostly the cheap layer.

Take a business handling 200 enquiries a week. You buy the AI in the CRM, because it’s there and it’s included. You buy a point tool for transcription, because somebody already built a better one than you’ll get. You wire the standard steps together with plain automation, because the rule never changes. Then the one custom piece reads the 40 enquiries that didn’t fit, pulls history from three systems that have never spoken, drafts the reply, and flags the 6 that need a person to decide.

One process, four technologies, and only the last slice was built. That’s what a sensible system looks like. AI agents vs chatbots vs automations has the detail on which layer does what.

If a proposal you’ve been sent is custom top to bottom, somebody hasn’t looked properly at your process. The reverse happens just as often, mind: businesses that buy 14 subscriptions and still have somebody spending Thursday afternoon fixing what the subscriptions got wrong. That afternoon was the part that needed building, and it’s been sitting there the whole time.

What happens when the answer is buy

We say so. Often.

An audit maps and costs every repeatable process you have and ranks them by what to fix first, and a fair share of what comes out of it is “turn this on in the software you already pay for” or “this is a £20 a month tool”. That’s a good outcome. It costs you nothing to run, it frees the budget for the one process where a build genuinely pays, and it’s a lot easier to trust the recommendation to build when the same person told you not to bother four times already.

The commercial side of that is simple enough. Whatever we do end up building, the price is fixed and agreed in writing before anything starts. No day rates, no scope-creep invoices, no paid discovery phase, and a full refund if you don’t approve the prototype. A supplier billing by the day has no reason to tell you the cheap answer, and the UK market makes that concrete: YunoJuno’s 2026 rates report puts the average developer day rate at £438, with the top 10% of contracts averaging £654. Multiply either by an open-ended number of days and you can see where the incentive points.

The short version

Buy the commodity. Own the thing that makes you different. Almost everything is the commodity.

Rent it when somebody already sells it, when it lives in one system, when the process is still moving, or when it costs you under a couple of hours a week. Build when the work crosses systems that don’t talk, when the exceptions are the expensive part, and when the volume has been steady long enough to be worth encoding.

Then count what it costs you by hand before you do either. That number is the only thing that tells you afterwards whether any of it worked.

Got a quote for a build and want a second opinion?

Send us the process. We’ll tell you whether it’s a commodity you can rent for the price of a couple of lunches, and we’ll say so plainly when it is.

If it isn’t, the price for what we build is fixed and agreed in writing before anything starts, with a full refund if you don’t approve the prototype.

Book a free call, or read what the AI audit covers first.

Want to see what this looks like for your business?

Book a free architecture audit. We'll map out where the bottlenecks are and what a custom platform could look like.

Book a free audit